CN104050517A

CN104050517A - Photovoltaic power generation forecasting method based on GRNN

Info

Publication number: CN104050517A
Application number: CN201410299370.XA
Authority: CN
Inventors: 董聪; 柳进; 刘广一; 于继来
Original assignee: Harbin Institute of Technology Shenzhen
Current assignee: Harbin Institute of Technology Shenzhen
Priority date: 2014-06-27
Filing date: 2014-06-27
Publication date: 2014-09-17

Abstract

基于GRNN神经网络的光伏发电预测方法，涉及一种基于广义回归(GRNN)神经网络光伏发电预测方法。所述方法首先考虑节气、天气、日照的因素，建立光伏曲线模式，提出基于广义回归神经网络光伏发电预测模型，并进行求解算法设计。本发明具有以下优点：GRNN网络的训练过程中不需要误差反向计算来修正权值，而只需要改变平滑参数σ来调节传递函数，减少了训练时间，加快网络学习速度；GRNN神经网络预测模型非线性映射能力强，逼近性能好，具有较强鲁棒性，适用于处理不稳定数据；GRNN神经网络的光伏发电预测技术，明显提高了预测精度；预测结果可为电网光电调度提供决策信息，对保证电网安全运行具有重要意义。A photovoltaic power generation prediction method based on a GRNN neural network relates to a photovoltaic power generation prediction method based on a generalized regression (GRNN) neural network. The method first considers solar terms, weather, and sunshine factors, establishes a photovoltaic curve model, proposes a photovoltaic power generation prediction model based on a generalized regression neural network, and designs a solution algorithm. The present invention has the following advantages: in the training process of the GRNN network, no reverse calculation of the error is required to modify the weight, but only the smoothing parameter σ is required to adjust the transfer function, which reduces the training time and speeds up the network learning; the GRNN neural network prediction model Strong nonlinear mapping ability, good approximation performance, strong robustness, suitable for dealing with unstable data; GRNN neural network photovoltaic power generation prediction technology, which significantly improves prediction accuracy; prediction results can provide decision-making information for power grid photoelectric dispatching, It is of great significance to ensure the safe operation of the power grid.

Description

Photovoltaic Power Generation Prediction Method Based on GRNN Neural Network

技术领域technical field

本发明涉及一种基于广义回归(GRNN)神经网络光伏发电预测方法。The invention relates to a method for predicting photovoltaic power generation based on a generalized regression (GRNN) neural network.

背景技术Background technique

由于太阳能等为可再生清洁能源，太阳能等新型能源在国内外得到了广泛的应用、太阳能的光伏发电功率与风力发电同样受天气、日照等不确定因素的影响，具有较强的随机间歇性，对电网安全运行带来了不稳定性。提高光伏发电功率预测精度，对发电计划制定，保证电网安全运行具有广泛的应用价值。Since solar energy is a renewable and clean energy source, new energy sources such as solar energy have been widely used at home and abroad. The photovoltaic power generation power of solar energy is also affected by uncertain factors such as weather and sunshine, and has strong random intermittency. It brings instability to the safe operation of the power grid. Improving the prediction accuracy of photovoltaic power generation has extensive application value for the formulation of power generation plans and ensuring the safe operation of the power grid.

发明内容Contents of the invention

本发明的目的是提供一种基于GRNN神经网络的光伏发电预测方法，首先考虑节气、天气、日照的因素，建立光伏曲线模式，提出基于广义回归(GRNN)神经网络光伏发电预测模型，并进行求解算法设计，目前国内尚未有相关技术的报道。The purpose of the present invention is to provide a photovoltaic power generation prediction method based on the GRNN neural network, first consider the factors of solar terms, weather, and sunshine, establish a photovoltaic curve model, propose a photovoltaic power generation prediction model based on the generalized regression (GRNN) neural network, and solve the problem Algorithm design, there is no report on related technologies in China at present.

本发明的目的是通过以下技术方案实现的：The purpose of the present invention is achieved through the following technical solutions:

一种基于GRNN神经网络的光伏发电预测方法，具体实施步骤如下：A photovoltaic power generation prediction method based on GRNN neural network, the specific implementation steps are as follows:

步骤1：分析光伏负荷的负荷特征和影响因子，运用K均值聚类方法对天气样本空间进行聚类分成二组，选取与预测日分为一组的各日光伏负荷数据建立负荷模式M＝{X₁,X₂,...,X_n+1}，其中选取与预测日最近的一日光伏负荷数据Y＝{y₁,y₂,…y_m}为网络期望输出，其余日光伏负荷数据M＝{X₁,X₂,...,X_n}为GRNN神经网络的输入向量，其中n为训练样本数，m为日光伏负荷数据个数(也是网络输出向量的维数)；Step 1: Analyze the load characteristics and influencing factors of photovoltaic load, use the K-means clustering method to cluster the weather sample space into two groups, and select the daily photovoltaic load data that is grouped with the forecast day to establish a load model M={ X ₁ ,X ₂ ,...,X _n+1 }, where the daily photovoltaic load data Y={y ₁ ,y ₂ ,…y _m } closest to the forecast date is selected as the expected output of the network, and the remaining daily photovoltaic load Data M={X ₁ ,X ₂ ,...,X _n } is the input vector of the GRNN neural network, where n is the number of training samples, and m is the number of daily photovoltaic load data (also the dimension of the network output vector);

步骤2：建立GRNN神经网络预测模型后，计算模式层神经元输出，并分别计算两类求和层神经元输出S_D和S_Nk，并计算出网络实际输出向量 $\hat{Y} = {{\hat{y}}_{1}, {\hat{y}}_{2}, . . ., {\hat{y}}_{m}} .$ Step 2: After establishing the GRNN neural network prediction model, calculate the neuron output of the pattern layer, and calculate the output _SD and S _Nk of the two types of summation layer neurons respectively, and calculate the actual output vector of the network $\hat{Y} = {{\hat{the y}}_{1}, {\hat{the y}}_{2}, . . ., {\hat{the y}}_{m}} .$

步骤3：对比网络实际输出和期望输出，计算误差目标函数，判断模型是否合格，如果合格进行步骤4，如不合格则通过再逐步增大平滑参数σ，修正模型之后重新计算模式层神经元输出、两类求和层神经元输出S_D和S_Nk和网络实际输出值直到误差函数小于给定精度，停止迭代，进行步骤4。Step 3: Compare the actual output of the network with the expected output, calculate the error objective function, and judge whether the model is qualified. If it is qualified, proceed to step 4. If it is not qualified, then gradually increase the smoothing parameter σ, and then recalculate the neuron output of the model layer after correcting the model , two types of summation layer neurons output S _D and S _Nk and the actual output value of the network Until the error function is less than the given precision, stop the iteration and go to step 4.

步骤4：将光伏负荷模式输入到GRNN神经网络预测模型中，按步骤3过程计算得到的网络输出即为光伏负荷预测向量 Step 4: Input the photovoltaic load pattern into the GRNN neural network forecasting model, and the network output calculated according to the step 3 process is the photovoltaic load forecasting vector

步骤5：对比预测结果与实际用户负荷，计算平均相对误差，VAR值等误差指标，并根据这些指标评价预测误差。Step 5: Comparing the forecast results with the actual user load, calculating the average relative error, VAR value and other error indicators, and evaluating the forecast error based on these indicators.

本发明具有以下优点：The present invention has the following advantages:

1、GRNN网络的训练过程中不需要误差反向计算来修正权值，而只需要改变平滑参数σ来调节传递函数，减少了训练时间，加快网络学习速度；1. During the training process of the GRNN network, there is no need to reverse the calculation of the error to correct the weight, but only need to change the smoothing parameter σ to adjust the transfer function, which reduces the training time and speeds up the network learning speed;

2、GRNN神经网络预测模型非线性映射能力强，逼近性能好，具有较强鲁棒性，适用于处理不稳定数据；2. The GRNN neural network prediction model has strong nonlinear mapping ability, good approximation performance, strong robustness, and is suitable for dealing with unstable data;

3、GRNN神经网络的光伏发电预测技术，明显提高了预测精度；3. The photovoltaic power generation prediction technology of GRNN neural network has significantly improved the prediction accuracy;

4、预测结果可为电网光电调度提供决策信息，对保证电网安全运行具有重要意义。4. The prediction results can provide decision-making information for power grid photoelectric dispatching, which is of great significance to ensure the safe operation of the power grid.

附图说明Description of drawings

图1为GRNN神经网络结构；Figure 1 is the GRNN neural network structure;

图2为GRNN神经网络预测算法流程。Figure 2 shows the flow of the GRNN neural network prediction algorithm.

具体实施方式Detailed ways

下面结合附图对本发明的技术方案作进一步的说明，但并不局限如此，凡是对本发明技术方案进行修改或者等同替换，而不脱离本发明技术方案的精神和范围，均应涵盖在本发明的保护范围中。The technical solution of the present invention will be further described below in conjunction with the accompanying drawings, but it is not limited to this. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered by the technical solution of the present invention. in the scope of protection.

本发明提供了一种基于GRNN神经网络的光伏发电预测方法，包括以下内容：The present invention provides a kind of photovoltaic power generation prediction method based on GRNN neural network, comprising the following contents:

一、按节气、天气、工作日和节假日等因素进行模糊聚类，选择与预测日相似的日光伏发电曲线，确定输入向量；建立预测GRNN神经网络结构；1. Carry out fuzzy clustering according to factors such as solar terms, weather, working days and holidays, select a daily photovoltaic power generation curve similar to the forecast day, and determine the input vector; establish a forecast GRNN neural network structure;

二、对预测GRNN神经网络结构进行训练，确定GRNN神经网络的光伏发电功率预测模型；2. Train the prediction GRNN neural network structure, and determine the photovoltaic power generation power prediction model of the GRNN neural network;

三、设计模型求解算法；3. Design the algorithm for solving the model;

四、确定平均相对误差和VAR值的误差评价指标，评价所提出方法的预测精度，分析预测误差风险。4. Determine the error evaluation index of average relative error and VAR value, evaluate the prediction accuracy of the proposed method, and analyze the risk of prediction error.

具体内容如下：The specific content is as follows:

(1)神经网络的结构(1) The structure of the neural network

GRNN神经网络是在RBF神经网络基础之上发展起来的典型的前馈型神经网络。典型的GRNN神经网络由四层组成：输入层、模式层(径向基层)、求和层、输出层。输入历史数据逐次通过四层运算得到网络预测输出，并根据输出的误差不断调节平滑参数σ，最终使误差最小。GRNN神经网络的结构如图1所示。GRNN neural network is a typical feed-forward neural network developed on the basis of RBF neural network. A typical GRNN neural network consists of four layers: input layer, pattern layer (radial base layer), summation layer, and output layer. The input historical data is successively obtained through four-layer operations to obtain the network prediction output, and the smoothing parameter σ is continuously adjusted according to the output error to minimize the error. The structure of GRNN neural network is shown in Figure 1.

网络的输入向量X_j＝[x_j1,x_j2,…,x_jm]^T，输出向量其中1≤j≤n，n为训练样本数，m代表网络输出向量维数。The input vector X _j of the network ＝[x _j1 ,x _j2 ,…,x _jm ] ^T , the output vector Where 1≤j≤n, n is the number of training samples, and m represents the dimension of the network output vector.

模式层神经元P_i的输出公式为：The output formula of the pattern layer neuron P _i is:

${p p}_{j j} = = exp exp [[- - \frac{{(({X x}_{j j} - - {X x}_{i i}))}^{T T} (({X x}_{j j} - - {X x}_{i i}))}{{22 σ σ}^{22}}]] - - - - - - ((11)) . .$

式中：X_j为网络输入的向量；X_i为神经元i对应的训练向量，j,i＝1,2,…,n。In the formula: X _j is the vector input by the network; _Xi is the training vector corresponding to neuron i, j,i=1,2,...,n.

网络的求和层的神经元分为两类：The neurons of the summation layer of the network are divided into two categories:

第一类神经元只有一个，即神经元S_D，这个神经元与模式层(神经元的连接权重固定为1，传递函数如下：There is only one neuron of the first type, that is, neuron S _D , this neuron and the pattern layer (the connection weight of the neuron is fixed at 1, and the transfer function is as follows:

${S S}_{D D.} = = {Σ Σ}_{j j = = 11}^{n no} exp exp [[- - \frac{{(({X x}_{j j} - - {X x}_{i i}))}^{T T} (({X x}_{j j} - - {X x}_{i i}))}{{22 σ σ}^{22}}]] - - - - - - ((22)) . .$

由公式可知，该函数功能为对模式层输出求和。It can be seen from the formula that the function of this function is to sum the output of the pattern layer.

第二类神经元为S_N1,S_N2,...,S_Nn，其数目与输出层神经元相同，传递函数功能为对模式层输出加权求和，模式层神经元P_j与求和层神经元S_Nk之间的传递函数如下：The second type of neurons are _S _N1 , S _N2 ,...,S _Nn , the number of which is the same as that of the neurons in the output layer. The transfer function between neurons S _Nk is as follows:

${S S}_{Nk Nk} = = {Σ Σ}_{j j = = 11}^{n no} {\overset{^^}{y the y}}_{k k} exp exp [[- - \frac{{(({X x}_{j j} - - {X x}_{i i}))}^{T T} (({X x}_{j j} - - {X x}_{i i}))}{{22 σ σ}^{22}}]] - - - - - - ((33)) . .$

公式中：为第k个网络输出值，1≤k≤m，m代表网络输出向量维数。网络的输出层神经元的数目和输出向量维数相同，每一个输出层神经元同第二类求和层神经元一一对应，只同求和层的S_D和对应的S_Nk两个神经元连接，第k个网络输出值为：formula: is the kth network output value, 1≤k≤m, m represents the dimension of the network output vector. The number of neurons in the output layer of the network is the same as the dimension of the output vector, and each neuron in the output layer corresponds to the neurons in the second summation layer one by one, only the _SD of the summation layer and the corresponding S _Nk two neurons Meta connection, the output value of the kth network is:

${\overset{^^}{y the y}}_{k k} = = \frac{{S S}_{Nk Nk}}{{S S}_{D D.}},, 11 \leq \leq k k \leq \leq m m - - - - - - ((44)) . .$

由上述分析可知，GRNN神经网络确定输入向量后，网络的权值已经确定无需修正，只要确定平滑参数σ就确定了预测模型。确定最佳平滑参数σ的过程就是网络的训练过程。训练方法为让平滑参数σ在一定的范围内逐级递增，不断比较目标函数，直到目标函数达到最小，将网络输出误差作为目标函数公式如下：It can be seen from the above analysis that after the GRNN neural network determines the input vector, the weight of the network has been determined without correction, and the prediction model is determined only by determining the smoothing parameter σ. The process of determining the optimal smoothing parameter σ is the training process of the network. The training method is to increase the smoothing parameter σ step by step within a certain range, and continuously compare the objective function until the objective function reaches the minimum. The network output error is used as the objective function formula as follows:

$min min ((E E.)) = = min min \frac{11}{n no} {Σ Σ}_{j j = = 11}^{n no} \frac{11}{22} {((\overset{^^}{Y Y} (({X x}_{j j})) - - Y Y (({X x}_{j j}))))}^{22} - - - - - - ((55)) . .$

(3)算法流程(3) Algorithm flow

基于K均值聚类的GRNN神经网络的电网光伏发电的预测方法，算法流程如图2所示。具体步骤如下：The prediction method of grid photovoltaic power generation based on K-means clustering GRNN neural network, the algorithm flow is shown in Figure 2. Specific steps are as follows:

步骤1：分析光伏负荷的负荷特征和影响因子，运用K均值聚类方法对天气样本空间进行聚类分成二组，选取与预测日分为一组的各日光伏负荷数据建立负荷模式M＝{X₁,X₂,...,X_n+1}，其中选取与预测日最近的一日光伏负荷数据Y＝{y₁,y₂,…y_m}为网络期望输出，其余日光伏负荷数据M＝{X₁,X₂,...,X_n}为GRNN神经网络的输入向量，其中n为训练样本数，m为日光伏负荷数据个数，也是网络输出向量的维数；Step 1: Analyze the load characteristics and influencing factors of photovoltaic load, use the K-means clustering method to cluster the weather sample space into two groups, and select the daily photovoltaic load data that is grouped with the forecast day to establish a load model M={ X ₁ ,X ₂ ,...,X _n+1 }, where the daily photovoltaic load data Y={y ₁ ,y ₂ ,…y _m } closest to the forecast date is selected as the expected output of the network, and the remaining daily photovoltaic load Data M={X ₁ ,X ₂ ,...,X _n } is the input vector of the GRNN neural network, where n is the number of training samples, m is the number of daily photovoltaic load data, and is also the dimension of the network output vector;

步骤2：建立GRNN神经网络预测模型后，计算模式层神经元P_i输出，并分别计算两类求和层神经元输出S_D和S_Nk，并计算出网络实际输出向量 $\hat{Y} = {{\hat{y}}_{1}, {\hat{y}}_{2}, . . ., {\hat{y}}_{m}};$ Step 2: After establishing the GRNN neural network prediction model, calculate the output of the pattern layer neuron P _i , and calculate the output _SD and S _Nk of the two types of summation layer neurons respectively, and calculate the actual output vector of the network $\hat{Y} = {{\hat{the y}}_{1}, {\hat{the y}}_{2}, . . ., {\hat{the y}}_{m}};$

步骤3：对比网络实际输出和期望输出，计算误差目标函数，判断模型是否合格，如果合格进行步骤4，如不合格则通过再逐步增大平滑参数σ，修正模型之后重新计算模式层神经元输出、两类求和层神经元输出S_D、S_Nk和网络实际输出值直到误差目标函数小于给定精度，停止迭代，进行步骤4；Step 3: Compare the actual output of the network with the expected output, calculate the error objective function, and judge whether the model is qualified. If it is qualified, proceed to step 4. If it is not qualified, then gradually increase the smoothing parameter σ, and then recalculate the neuron output of the model layer after correcting the model , two types of summation layer neurons output S _D , S _Nk and the actual output value of the network Until the error objective function is less than the given accuracy, stop the iteration and go to step 4;

步骤5：对比预测结果与实际用户负荷，计算误差指标，并根据误差指标评价预测误差。Step 5: Comparing the forecast result with the actual user load, calculating the error index, and evaluating the prediction error according to the error index.

Claims

1. the photovoltaic generation Forecasting Methodology based on GRNN neural network, is characterized in that described Forecasting Methodology is:

Step 1: analyze load characteristic and the factor of influence of photovoltaic load, use K means clustering method to carry out cluster to weather sample space and be divided into two groups, choose and predict that being divided into each day photovoltaic load data of a group day sets up load pattern M={X ₁, X ₂..., X _n+1, wherein choose and predict a day nearest one day photovoltaic load data Y={ _y1, _y2, _ymbe network desired output, all the other day photovoltaic load data M={X ₁, X ₂..., X _nbe the input vector of GRNN neural network, and wherein n is number of training, m is the dimension of network output vector;

Step 2: set up after GRNN neural network prediction model computation schema layer neuron P _ioutput, and calculate respectively two class summation layer neuron output S _dand S _nk, and calculate the actual output vector of network

\hat{Y} = {{\hat{y}}_{1}, {\hat{y}}_{2}, . . ., {\hat{y}}_{m}};

Step 3: the contrast actual output of network and desired output, error of calculation objective function, whether judgment models is qualified, if qualified, carry out step 4, as defective, by progressively increasing again smoothing parameter σ, after correction model, recalculate the output of mode layer neuron, two class summation layer neuron output S _d, S _nkwith network real output value until error objective function is less than given accuracy, stop iteration, carry out step 4;

Step 4: photovoltaic load pattern is input in GRNN neural network prediction model, and the network output obtaining by step 3 process computation is photovoltaic load prediction vector

Step 5: contrast predicts the outcome and actual user's load, error of calculation index, and evaluate predicated error according to error criterion.

2. the photovoltaic generation Forecasting Methodology based on GRNN neural network according to claim 1, is characterized in that described mode layer neuron P _ioutput formula be:

p_{j} = \exp [- \frac{{(X_{j} - X_{i})}^{T} (X_{j} - X_{i})}{{2 σ}^{2}}],

In formula: X _jinput vector for network; X _ifor training vector corresponding to neuron i, j, i=1,2 ..., n, n is number of training.

3. the photovoltaic generation Forecasting Methodology based on GRNN neural network according to claim 1, is characterized in that described neuron S _dtransport function is as follows:

S_{D} = Σ_{j = 1}^{n} \exp [- \frac{{(X_{j} - X_{i})}^{T} (X_{j} - X_{i})}{{2 σ}^{2}}],

In formula: X _jinput vector for network; X _ifor training vector corresponding to neuron i, i=1,2 ..., n, n is number of training.

4. the photovoltaic generation Forecasting Methodology based on GRNN neural network according to claim 1 and 2, is characterized in that described mode layer neuron P _iwith summation layer neuron S _nkbetween transport function as follows:

S_{Nk} = Σ_{j = 1}^{n} {\hat{y}}_{k} \exp [- \frac{{(X_{j} - X_{i})}^{T} (X_{j} - X_{i})}{{2 σ}^{2}}],

In formula: be k network output valve, 1≤k≤m, m represents network output vector dimension.

5. the photovoltaic generation Forecasting Methodology based on GRNN neural network according to claim 1, is characterized in that the number of described network output is identical with output vector dimension, and each output element is corresponding one by one with Equations of The Second Kind summation layer neuron, with the S of summation layer _dwith corresponding S _nktwo neurons connect, and output valve is:

{\hat{y}}_{k} = \frac{S_{Nk}}{S_{D}}, 1 \leq k \leq m,

M represents network output vector dimension.

6. the photovoltaic generation Forecasting Methodology based on GRNN neural network according to claim 1, is characterized in that described error objective function is as follows:

\min (E) = \min \frac{1}{n} Σ_{j = 1}^{n} \frac{1}{2} {(\hat{Y} (X_{j}) - Y (X_{j}))}^{2},

In formula: X _jinput vector for network; for being input as X _jprerequisite under, the actual output vector of network; Y is network desired output vector; N is number of training.